Abstract Body: Introduction: Several pre-hospital scales predict ischemic stroke due to large vessel occlusion; however, these scales often fail to detect atypical endovascular thrombectomy (EVT) cases, including posterior circulation and distal vessel occlusion. Additionally, these pre-hospital scales cannot differentiate between stroke subtypes, such as ischemic stroke and intracerebral hemorrhage. We aimed to develop machine learning (ML) models to predict stroke subtypes and EVT eligibility.
Methods: We conducted an analysis using data from the Japan Stroke Data Bank, a nationwide acute stroke registry. Patients with ischemic stroke, intracerebral hemorrhage, or subarachnoid hemorrhage who were hospitalized between 2016 and 2020 were included in this study. We developed two ML models: (1) a model to predict patients who received EVT, and (2) to predict stroke categories (ischemic stroke treated with EVT, ischemic stroke treated with tPA, ischemic stroke without EVT or tPA, intracerebral hemorrhage, and subarachnoid hemorrhage). Patient data were divided into the derivation cohort (data from 2016 to 2019) and the validation cohort (data in 2020). The input variables for the ML models consisted of 129 clinical characteristics, including past medical history and neurological examination findings. Imaging and laboratory data were not utilized in the ML process. We initially developed the ML models using LightGBM algorithm with all 129 variables. Subsequently, we selected the top 10 variables for model development by the feature importance and Brute-force method, comparing commonly used pre-hospital scales to predict patients with EVT.
Results: Of the 62,588 patients included in the study, 4,353 were treated with EVT, 3,001 with tPA, 39,710 had ischemic stroke treated without EVT or tPA, 12,498 had intracerebral hemorrhage, and 3,028 had subarachnoid hemorrhage. The area under the curve (AUC) for predicting patients treated with EVT was 0.828 (95% CI 0.817-0.839). For predicting the five stroke categories, the multiclass AUC was 0.848, and the micro-average F1 score was 0.705. Using the brute-force algorithm to select 10 variables, the ML model for predicting EVT treatment achieved an AUC of 0.794 (95% CI 0.784-0.805), outperforming commonly used pre-hospital scales (AUC 0.584-0.690).
Conclusions: The ML models utilizing only clinical characteristics can accurately predict stroke subtypes and EVT eligibility, offering a potential improvement over current pre-hospital scales.